This is the code and data from my 2016 Journal of New Music Research paper. If you use it, please cite it:

@article{McLeod:16,
  author  = {McLeod, Andrew and Steedman, Mark},
  title   = {HMM-Based Voice Separation of MIDI Performance},
  journal = {Journal of New Music Research},
  year    = {2016},
  volume  = {45},
  number  = {1},
  pages   = {17--26}
}

The goal of this project is to have a Java program which is able to split MIDI performance data into monophonic voices.

This command will write the resulting voices of the given FILES into new MIDI files in the voice directory, with default settings.

java -cp bin voicesplitting.voice.hmm.HmmVoiceSplittingModelTester -w voice FILES

• Add -l if the MIDI files are live performance (unquantized) MIDI.
• If the computation takes a long time for some files, you can use the -M INT flag to set the maximum number of voices to output (e.g., -M 10).

Note that some MIDI files separate voices by channel, while some do so by track. This project uses channel by default, but this can be changed by using the -T flag when running from the command line, or by unchecking the Use Channel box (and then reloading the file) if using the GUI. MIDI files written out by this software will have voices divided by both channel and track.

You are free to use this software as is allowed by the MIT License. I only ask that you please cite it as my work where appropriate, including the paper on my website which has been published in the Journal of New Music Research.

This document contains some basic code examples and a general overview of how to use the classes in this project. All specific documentation for the code found in this project can be found in the Javadocs.

An in-depth walkthrough of how to use the GUI can be found in the GUI Readme.

The java files can all be compiled into class files in a bin directory using the Makefile included with the project with the following command: make.

Once the class files are installed in the bin directory, the project can be run, either using the GUI or from the command line.

To run the GUI version of the project, use the command java -cp bin voicesplitting.gui.VoiceSplittingGUI. An in-depth walkthrough of how to use the GUI can be found in the GUI Readme.

To run the project from the command line, use the command

java -cp bin voicesplitting.voice.hmm.HmmVoiceSplittingModelTester ARGS Files

Here, Files should be a list of 1 or more MIDI files or directories containing only MIDI files. Any directories listed will be searched recursively for all files, and if any of the files found are not in MIDI format, an exception will be thrown.

To write results out as tracks to a new MIDI file:

java -cp bin voicesplitting.voice.hmm.HmmVoiceSplittingTester -w voice FILES (also add -l if the files are live performance).

This will write all of the files out to new MIDI files in the directory voice.

ARGS: One of these is required:

• -r = Run a test (if used with -t, we will use the tuned parameters instead of any given).
• -t [STEPS] = Tune to maximize F1-Measure, and optionally set the number of steps to make within each parameter range to an Integer value (default = 5). It is HIGHLY recommended to use this training method rather than your own script because it runs the tests in parallel as much as possible to speed up training.
• -e = Extract and print the separated voices.
• -w DIR = Write the separated voices out to a new MIDI files in the DIR directory.

Optional arguments:

• -v = Verbose (print out each song and each individual voice when running).
• -T = Use tracks as correct voice (instead of channels). See Troubleshooting for more information.

If running with -r, -e, or -w, the following arguments can be used to change the parameter settings from their default values (those with which we tested the computer generated WTC fugues in the paper):

• -l = Set the parameters to those best for live performance.
• -b INT = Set the Beam Size parameter to the value INT.
• -n DOUBLE = Set the New Voice Probability parameter to the value DOUBLE.
• -h INT = Set the Pitch History Length parameter to the value INT.
• -g DOUBLE = Set the Gap Std Micros parameter to the value DOUBLE.
• -p DOUBLE = Set the Pitch Std parameter to the value DOUBLE.
• -m DOUBLE = Set the Min Gap Score parameter to the value DOUBLE.
• -M INT = Set the maximum number of voices that the program will generate (default unlimited). Can speed up computation in some cases.

The -e output has the following format:

songID noteID voiceID onsetTime(microseconds) offsetTime(microseconds) pitch velocity

The standard (non-verbose) -r output has the following format:

(b,n,h,g,p,m) = V=Average Voice Consistency P=Precision R=Recall F1=F1-Measure

Here, the letters within the parentheses on the left represent the settings of the parameters as listed in the Arguments section above. The values on the right correspond to the Average Voice Consistency, Precision, Recall, and F1-Measure, each as an average over all of the songs given in the Files argument. These metrics are defined precicely in the paper.

The verbose -r output has the standard output as its final line, and also prints the following for each of the songs given in the Files argument:

FileName Correct / Total = Voice Consistency ... P=Precision R=Recall F1=F1-Measure

Here, the "Correct / Total = Voice Consistency" line is repeated per voice, and the Voice Consistency metric is again defined in the paper.

The training output consists of the standard (or verbose) -r output for each unique setting of the parameters, followed by a single line of output of BEST = followed by the non-verbose output from the parameter settings which achieved the greatest F1-Measure.

To tune on the files in the directory "midi" with a step of 7 and print verbose results:

java -cp bin voicesplitting.voice.hmm.HmmVoiceSplittingModelTester -t 7 -v midi

To test on the files in the directory "midi" using all default values except a beam size of 25:

java -cp bin voicesplitting.voice.hmm.HmmVoiceSplittingModelTester -r -b 25 midi

To test on the files in the directory "midi" using the default parameter settings:

java -cp bin voicesplitting.voice.hmm.HmmVoiceSplittingModelTester -r midi

Most Exceptions that occur while running this program should print a useful error message out to stderr. In general, there are 2 main types of errors:

• I/O Errors: These indicate that some I/O error occured while trying to read a MIDI file. These could indicate that the program does not have read permission on the given file, or that the file does not exist. Check for typos in the file names.
• MIDI errors: These indicate that the MIDI files given are not in a format that can be read by Java. There is currently no way to solve this issue besides looking for new MIDI files.

If the results you are getting seem very low, try using the -T flag if running from the command line, or unchecking the Use Channel box (and then reloading the file) if using the GUI. Some MIDI files divide notes into voices by track while others do it by channel. MIDI files created by this software does both.

Please let me know if you are interested in using my work. If you run into any problems installing it, using it, extending it, or you'd like to see me add any additional features, please let me know either by email or by submitting an issue on github. Any and all questions are always welcome.

There is a paper which has been published in the Journal of New Music Research available on my website with further documentation. Please cite this if you use my code or the paper.

Thanks, and enjoy! Andrew McLeod